WEBVTT
1
00:00:14.200 --> 00:00:17.160
Good morning, everyone. My name is Erie Tedeschi.
2
00:00:17.400 --> 00:00:20.800
I'm the head of economic insights and research at Stripe.
3
00:00:21.760 --> 00:00:25.180
There has been no shortage of excitement, investment,
4
00:00:25.760 --> 00:00:28.500
and claims about artificial intelligence.
5
00:00:29.300 --> 00:00:33.820
Artificial intelligence came from research labs to our work and our
6
00:00:33.880 --> 00:00:37.520
lives probably faster than any technology in history.
7
00:00:38.260 --> 00:00:40.320
But because it's such a young technology,
8
00:00:40.920 --> 00:00:45.260
that also brings with it so much uncertainty about what its actual economic
9
00:00:45.300 --> 00:00:48.400
effects are. So what we've seen already,
10
00:00:49.180 --> 00:00:53.420
what is coming down the pipeline, and what this all means for growth,
11
00:00:53.800 --> 00:00:55.720
productivity, the labor market.
12
00:00:56.280 --> 00:00:58.520
That's what we're going to talk about in our panel today,
13
00:00:58.580 --> 00:01:01.460
and I am so excited to welcome this panel.
14
00:01:05.820 --> 00:01:10.080
First of all, Peter McCrory is the head of economics at Anthropic.
15
00:01:10.760 --> 00:01:14.600
He sees this from inside a leading artificial intelligence lab,
16
00:01:15.240 --> 00:01:17.840
where the capabilities are being built. Ara Kharazian
17
00:01:27.080 --> 00:01:31.800
is an economist at Ramp. He sees the impact in real-time spending data,
18
00:01:31.940 --> 00:01:36.140
so how companies are allocating the dollars for artificial intelligence. Ara,
19
00:01:36.200 --> 00:01:37.033
thank you for joining.
20
00:01:41.540 --> 00:01:42.560
And then last but not least,
21
00:01:42.800 --> 00:01:46.760
Basil Halperin is an assistant professor of economics at the University of
22
00:01:46.800 --> 00:01:51.480
Virginia. So Basil brings the academic lens on growth, labor economics,
23
00:01:51.600 --> 00:01:54.800
and what history tells us about technology transitions.
24
00:01:59.360 --> 00:02:01.380
Okay. First of all, thank you all for being here.
25
00:02:01.440 --> 00:02:05.140
I'm really excited about this panel. Before we get to the macro questions,
26
00:02:05.200 --> 00:02:08.900
and we will get to macro questions, let's start with a personal question. So
27
00:02:10.640 --> 00:02:12.560
Peter, I'll start with you and we'll go down the line.
28
00:02:12.800 --> 00:02:17.600
What's one thing that you now do with AI that you can't imagine not doing with
29
00:02:17.640 --> 00:02:18.473
AI anymore?
30
00:02:19.440 --> 00:02:21.980
It's a great question. I mean, in some sense,
31
00:02:23.160 --> 00:02:28.120
what Claude opens up is this ability to do so much more than what I as an
32
00:02:28.180 --> 00:02:29.960
economist have training to do.
33
00:02:30.320 --> 00:02:35.040
So one concrete example is I can now build interactive dashboards to visualize
34
00:02:35.100 --> 00:02:39.740
the data in almost instantaneously and not only explore that for myself,
35
00:02:39.860 --> 00:02:44.760
but send that dashboard to colleagues and illustrate the insights
36
00:02:44.780 --> 00:02:49.620
that my team is finding and showcase it or even cast a
37
00:02:49.680 --> 00:02:53.080
vision of what we might want to build together. As an economist,
38
00:02:53.720 --> 00:02:58.220
one thing that I feel already is that the technology helps me
39
00:02:58.700 --> 00:03:02.920
iterate and explore ideas much faster than I could have before.
40
00:03:03.420 --> 00:03:07.640
You can send Claude off on a journey to estimate a regression or do some
41
00:03:07.660 --> 00:03:11.460
statistical analysis to find out if the idea is actually worth exploring
42
00:03:11.860 --> 00:03:12.693
further.
43
00:03:14.000 --> 00:03:14.833
Ara?
44
00:03:14.900 --> 00:03:15.733
I completely agree.
45
00:03:15.740 --> 00:03:20.620
It's this incredibly powerful RA that allows you to iterate and try
46
00:03:20.680 --> 00:03:24.360
out new ideas extremely quickly, swap out different methodologies to see if,
47
00:03:24.440 --> 00:03:26.420
hey, you may have found this one result somewhere,
48
00:03:26.480 --> 00:03:30.200
but if you change a couple assumptions, you get something completely different.
49
00:03:30.480 --> 00:03:32.460
So I think it makes everyone's research a little bit stronger.
50
00:03:34.200 --> 00:03:37.360
Yeah. So besides coding, which I think everyone's aware of,
51
00:03:37.520 --> 00:03:41.780
I think the effect of AI on the ability to do math is kind of
52
00:03:42.660 --> 00:03:45.100
growing and underappreciated. So as an economist,
53
00:03:45.160 --> 00:03:47.580
I spend a lot of the day pushing Greek letters around the whiteboard,
54
00:03:48.060 --> 00:03:51.220
and I drop a lot of negative signs because I'm not a full-time mathematician,
55
00:03:51.400 --> 00:03:55.060
things like that. And AI really helps just do the math for me.
56
00:03:55.800 --> 00:04:00.200
Wonderful. So now let's get into the macro questions because I can't resist.
57
00:04:00.760 --> 00:04:03.800
So we keep hearing about how AI is changing work,
58
00:04:04.480 --> 00:04:08.160
but we don't really see that in the productivity data yet. So I looked this up.
59
00:04:08.700 --> 00:04:11.080
Productivity between 2024 and 2025,
60
00:04:11.340 --> 00:04:15.560
if you look at business sector output per worker grew 2.2%,
61
00:04:16.500 --> 00:04:20.660
which is fine, solid, maybe even a little bit more than solid,
62
00:04:21.060 --> 00:04:23.960
but not exactly gangbusters. In 1999,
63
00:04:24.560 --> 00:04:26.460
in the midst of the dot-com boom,
64
00:04:26.940 --> 00:04:30.780
business output per worker grew by 4% that year.
65
00:04:31.360 --> 00:04:34.900
If you look at total factor productivity, which is harder to measure,
66
00:04:35.000 --> 00:04:38.460
but that's really what economists think of when we think of technological
67
00:04:38.500 --> 00:04:42.120
growth, that's the efficiency with which we use labor and capital.
68
00:04:42.380 --> 00:04:46.840
That grew 0.8% in 2025 versus roughly 2%
69
00:04:47.500 --> 00:04:51.000
in 1999. So even more mid,
70
00:04:51.820 --> 00:04:56.380
not quite as impressive. So what's going on here? Is it just too early?
71
00:04:56.600 --> 00:05:01.480
Is this the solo paradox in action or is there something unique about AI
72
00:05:01.860 --> 00:05:05.380
where it just takes a long time for it to show up in the productivity data?
73
00:05:05.760 --> 00:05:07.480
Basil, why don't I start with you and we'll come back this way?
74
00:05:07.780 --> 00:05:10.320
Yeah. I think John's excellent talk this morning,
75
00:05:10.440 --> 00:05:14.300
still a lot of our talking points here where there really is just this slow
76
00:05:14.440 --> 00:05:18.280
diffusion through the economy where firms need to reorganize their internal
77
00:05:18.340 --> 00:05:20.680
processes. So like if you think, for example,
78
00:05:21.190 --> 00:05:25.300
1982 Commodore desktop comes out, five years later,
79
00:05:25.440 --> 00:05:29.320
there's this famous solo clip. We see the computer age everywhere,
80
00:05:29.380 --> 00:05:32.860
but in the productivity statistics, took another 10 years for the dot-com boom.
81
00:05:33.580 --> 00:05:36.260
And the dot-com boom is really driven by capital investment,
82
00:05:38.360 --> 00:05:40.160
will be the analog of building data centers today.
83
00:05:40.280 --> 00:05:44.220
And then it wasn't even until another five years later in the early 2000s that
84
00:05:44.560 --> 00:05:47.420
finally TFP was growing at its fastest rate.
85
00:05:47.780 --> 00:05:52.280
So even if things are growing five times as fast today as they were during the
86
00:05:52.340 --> 00:05:55.140
dot-com era, the computer age, that's like
87
00:05:57.380 --> 00:06:02.260
four years until from the desktop computer
88
00:06:02.300 --> 00:06:03.580
until we see things in TFP.
89
00:06:06.060 --> 00:06:07.900
I disagree. I think we are seeing some of the results.
90
00:06:08.540 --> 00:06:11.420
I don't think that you're going to see it in aggregate statistics that are
91
00:06:11.520 --> 00:06:15.740
weighed down by the vast majority of businesses that can't take advantage of AI
92
00:06:15.860 --> 00:06:16.640
effectively, right?
93
00:06:16.640 --> 00:06:21.160
A doctor's office or a dentist's office is not going to grow dramatically more
94
00:06:21.200 --> 00:06:25.440
because it has AI. And you might not even see it in some sector level data,
95
00:06:27.260 --> 00:06:29.180
even for the tech sector or the finance sector,
96
00:06:29.460 --> 00:06:31.940
two sectors that we know are pretty far ahead in AI adoption.
97
00:06:32.100 --> 00:06:35.440
But within those sectors, there are stories, more than just stories,
98
00:06:36.060 --> 00:06:40.860
of firms that are growing their revenues faster without
99
00:06:40.920 --> 00:06:44.780
growing their headcount. And then if you talk to the CFOs of those companies,
100
00:06:45.920 --> 00:06:50.800
they explicitly say that there has been a decoupling between revenue growth and
101
00:06:50.840 --> 00:06:51.673
headcount growth.
102
00:06:52.340 --> 00:06:55.360
And that's only at the companies that are pretty far along ahead in the AI
103
00:06:55.380 --> 00:06:59.120
adoption curve.
And that's really hard to identify in a data set when you're
104
00:06:59.140 --> 00:07:00.100
just looking at the aggregate.
105
00:07:02.180 --> 00:07:06.220
So I tend to kind of agree with both of these perspectives,
106
00:07:06.280 --> 00:07:11.140
which is it is undeniable that it's a general purpose technology that is poised
107
00:07:11.160 --> 00:07:15.980
to have very large and transformative effects on almost every sector and every
108
00:07:16.100 --> 00:07:19.880
occupation to some extent will be transformed by the technology.
109
00:07:20.820 --> 00:07:24.060
But then there's this question of how does that theoretical capability meet the
110
00:07:24.120 --> 00:07:24.953
real world?
111
00:07:25.260 --> 00:07:28.860
And so we do this exercise in a report that we put out in early March where we
112
00:07:28.960 --> 00:07:33.920
compare theoretical capabilities of large language models versus where Claude is
113
00:07:33.980 --> 00:07:37.080
actually used in automated ways for work purposes.
114
00:07:37.420 --> 00:07:42.280
And there's a huge gap for computer and mathematical occupations like software
115
00:07:42.300 --> 00:07:45.400
engineers, that theoretical exposure is close to 95%,
116
00:07:45.980 --> 00:07:50.880
but we only see about 34% of those tasks showing up in our
117
00:07:50.920 --> 00:07:52.540
data. And in general,
118
00:07:53.500 --> 00:07:57.820
it is the case that we see signs of adoption moving much faster than previous
119
00:07:57.900 --> 00:08:02.200
waves of transformative technologies.
About a year ago when we launched our
120
00:08:02.300 --> 00:08:03.320
economic index,
121
00:08:04.120 --> 00:08:08.800
we estimated that about 36% of jobs had at least a quarter of the tasks typical
122
00:08:08.880 --> 00:08:10.960
of those jobs showing up in our data.
123
00:08:11.320 --> 00:08:13.680
That number in just a year has risen to 50%.
124
00:08:14.040 --> 00:08:18.320
So one out of every two jobs for the US economy has a nontrivial share of the
125
00:08:18.400 --> 00:08:23.360
activity that you do in that job showing up in our data. Across the US,
126
00:08:23.880 --> 00:08:28.400
we're seeing rapid diffusion of Claude across states moving
127
00:08:28.720 --> 00:08:33.680
arguably five to 10 times faster than this benchmark of a prominent study
128
00:08:33.820 --> 00:08:34.740
that came out last year,
129
00:08:36.160 --> 00:08:41.160
documenting how long it typically takes transformative technologies to spread
130
00:08:41.200 --> 00:08:44.240
across the US. So I think we're actually moving quite fast,
131
00:08:44.480 --> 00:08:47.160
but it will take some time to show up in the aggregate statistics.
132
00:08:47.600 --> 00:08:50.400
And if it doesn't show up in the aggregate statistics,
133
00:08:50.520 --> 00:08:53.160
but we do see signs that say in the sectoral statistics,
134
00:08:53.960 --> 00:08:58.280
should that bother us or is that just going to be sort of the way that AI works?
135
00:08:58.360 --> 00:09:00.560
And I'll open this up to all three of you, no order.
136
00:09:01.800 --> 00:09:02.040
Well,
137
00:09:02.040 --> 00:09:04.640
that's where you get this situation where there are some companies that are
138
00:09:04.680 --> 00:09:09.600
growing significantly faster than other firms and traditional economic intuition
139
00:09:09.680 --> 00:09:12.920
would tell us that more firms are then going to enter those sectors where there
140
00:09:12.960 --> 00:09:16.120
are profits to be made and more workers might enter those sectors where there
141
00:09:16.120 --> 00:09:17.000
are profits to be made.
142
00:09:17.640 --> 00:09:21.320
Traditional economic thinking would also tell us that it's really hard to do
143
00:09:21.440 --> 00:09:26.040
those transitions because people study up for one thing or
144
00:09:27.080 --> 00:09:31.080
just making the transition between industries is often very difficult both for
145
00:09:31.120 --> 00:09:36.080
workers and for enterprise. That transition is often easier to make for AI.
146
00:09:36.640 --> 00:09:39.480
You can study up on it fairly quickly.
147
00:09:39.640 --> 00:09:44.320
The technology is designed to allow people to get acquainted with it fairly
148
00:09:44.400 --> 00:09:46.200
quickly. And so a lot of our
149
00:09:48.520 --> 00:09:51.560
traditional methods of thinking about the constraints of the transitions,
150
00:09:52.280 --> 00:09:53.320
I think fall apart here.
151
00:09:54.160 --> 00:09:59.040
I guess I would be very surprised if over the next decade we do not see clear
152
00:09:59.280 --> 00:10:02.120
markers of aggregate lift and productivity.
153
00:10:03.040 --> 00:10:08.020
So one of the exercises that we do in the Anthropic Economic Index is
154
00:10:08.120 --> 00:10:12.340
we have Claude look at the conversations that people are having in a
155
00:10:12.380 --> 00:10:16.700
privacy-preserving way and estimate what task are they doing and how much time
156
00:10:16.740 --> 00:10:18.180
did they save doing that task?
157
00:10:18.640 --> 00:10:23.580
Compiling information from reports might take people 10 hours to read all
158
00:10:23.640 --> 00:10:27.560
the documents and then a few more hours to put together a one-page memo.
159
00:10:28.100 --> 00:10:30.480
Claude can do that sort of thing in 15 to 20 minutes.
160
00:10:31.000 --> 00:10:34.880
When you add up the efficiency gains at the task level using standard macro
161
00:10:34.980 --> 00:10:38.220
growth accounting, for those that are familiar with this literature,
162
00:10:38.300 --> 00:10:39.800
it's referred to as Hulten's Theorem,
163
00:10:40.440 --> 00:10:45.160
you generate a number of 1.8-percentage-point increase in labor productivity
164
00:10:45.260 --> 00:10:47.480
growth each year over the next decade,
165
00:10:47.840 --> 00:10:52.620
if that's how long it takes for the diffusion process to unfold.
That's just
166
00:10:52.660 --> 00:10:55.320
current models and current usage patterns.
167
00:10:55.620 --> 00:11:00.120
As businesses figure out how to restructure operations that could amplify the
168
00:11:00.140 --> 00:11:03.060
effect, these models are improving very fast,
169
00:11:03.460 --> 00:11:07.460
and so that's another force pushing in the direction of large productivity
170
00:11:07.520 --> 00:11:12.240
gains, not from the capital expenditure needed to build out for AI,
171
00:11:12.340 --> 00:11:16.280
but from the actual efficiency gain that the technology brings.
172
00:11:17.860 --> 00:11:21.840
I'm as bullish, I think, as anyone in this room on the potential effects of AI,
173
00:11:21.980 --> 00:11:26.860
but I do want to caution about anecdotes where there is, for example,
174
00:11:26.920 --> 00:11:31.700
this famous study from the organization METR where they let a group of software
175
00:11:31.720 --> 00:11:36.460
engineers use AI for some coding projects, but not for others.
176
00:11:36.900 --> 00:11:38.880
And afterwards, they asked,
177
00:11:39.120 --> 00:11:43.280
"How much do you think AI sped you up on those projects?" And compared that to
178
00:11:43.460 --> 00:11:47.140
how much did it actually speed them up on average? And of course,
179
00:11:47.200 --> 00:11:49.540
everyone predicted that it sped them up 20% or something,
180
00:11:49.840 --> 00:11:51.460
but instead it made them marginally slower.
181
00:11:52.520 --> 00:11:54.080
It's hard to estimate your own productivity,
182
00:11:54.560 --> 00:11:56.600
estimating the aggregate productivity is even harder.
183
00:11:56.880 --> 00:12:00.760
So, I'm going to throw out a few topics. We'll start with Peter.
184
00:12:00.820 --> 00:12:04.500
We'll go on down the line for each one. For each one,
185
00:12:04.600 --> 00:12:07.480
I want you to tell me if it's overrated or underrated.
186
00:12:07.560 --> 00:12:09.760
And since we're all two-handed economists here.
187
00:12:10.120 --> 00:12:10.460
Totally.
188
00:12:10.460 --> 00:12:13.620
Fine if you need to add some nuance and if you want to talk about things more,
189
00:12:14.340 --> 00:12:17.460
totally understand. We did advertise that these are economists talking,
190
00:12:17.540 --> 00:12:20.840
so none of you should be surprised by how lengthy some of our answers are.
191
00:12:22.740 --> 00:12:27.400
Okay. Let's start with current AI CapEx sustainability in the near term.
192
00:12:27.660 --> 00:12:29.920
Peter, overrated or underrated in our discussions.
193
00:12:31.100 --> 00:12:33.160
I think the demand for intelligence is
194
00:12:35.200 --> 00:12:38.180
maybe insatiable, at least for the foreseeable future.
195
00:12:38.500 --> 00:12:42.220
And so it's unclear if we're... Yeah,
196
00:12:42.300 --> 00:12:46.780
there's going to be a lot of demand and we're maybe even running into supply
197
00:12:46.840 --> 00:12:47.540
constraints already.
198
00:12:47.540 --> 00:12:50.440
So I don't know which direction that puts overrated or underrated,
199
00:12:50.500 --> 00:12:51.420
but I'll let-.
200
00:12:52.040 --> 00:12:53.480
It's important. Yes.
201
00:12:54.280 --> 00:12:55.540
Ara? Properly rated.
202
00:12:55.740 --> 00:12:58.660
I think that there's a reasonable concern that some companies are spending too
203
00:12:58.680 --> 00:12:59.513
much. At the same time,
204
00:12:59.560 --> 00:13:02.100
these companies are making a lot of money selling this technology.
205
00:13:02.320 --> 00:13:04.460
As long as that continues, I think it's generally going to be fine.
206
00:13:06.300 --> 00:13:10.900
So I think bubble concerns wildly overrated or somewhat
207
00:13:10.920 --> 00:13:13.720
overrated, but sustainability of CapEx,
208
00:13:14.140 --> 00:13:16.660
we have to appreciate how fast CapEx has been growing.
209
00:13:16.720 --> 00:13:20.420
It's like over 3X per year, that's like 10X every two years.
210
00:13:20.740 --> 00:13:25.260
We basically start to run out of GDP, 2030, 2032. AI CapEx,
211
00:13:25.400 --> 00:13:28.780
if you kept growing at that rate is 10, 30% of the economy.
212
00:13:29.060 --> 00:13:34.020
You just literally run out of GDP unless GDP starts 10x'ing every two years,
213
00:13:34.080 --> 00:13:35.300
which would be of course crazy.
214
00:13:35.680 --> 00:13:36.513
That'd be great.
215
00:13:38.380 --> 00:13:42.520
Agents going mainstream in the Fortune 500 in the short term. Peter.
216
00:13:44.900 --> 00:13:49.180
I guess maybe I would say underrated in the sense that these models are
217
00:13:49.200 --> 00:13:54.020
increasingly capable of handling complex tasks and as they
218
00:13:54.060 --> 00:13:55.120
become more reliable,
219
00:13:55.740 --> 00:13:59.840
delegating is the very clear future.
220
00:14:01.560 --> 00:14:02.393
Underrated.
221
00:14:03.700 --> 00:14:07.520
Slightly overrated. I don't find agents that useful,
222
00:14:07.580 --> 00:14:09.580
but I'm a pointy-headed economist.
223
00:14:12.780 --> 00:14:16.980
AI-driven large-scale white-collar employment crisis. Peter,
224
00:14:17.160 --> 00:14:18.180
overrated or underrated.
225
00:14:21.160 --> 00:14:23.920
I would say it's not yet in the data.
226
00:14:24.860 --> 00:14:29.180
I have huge error bars over the future. Typically,
227
00:14:30.460 --> 00:14:33.760
new technologies automate some things, but create new tasks.
228
00:14:34.200 --> 00:14:39.160
Can we create new tasks fast enough to provide meaningful opportunity for
229
00:14:39.200 --> 00:14:40.033
displaced workers?
230
00:14:41.300 --> 00:14:44.220
Overrated for most jobs, properly rated for tech jobs.
231
00:14:45.620 --> 00:14:47.840
Maybe overrated for all.
232
00:14:48.200 --> 00:14:52.620
So I think of the case of translators perhaps in the last few years or call
233
00:14:52.680 --> 00:14:55.300
center workers where you might have really imagined that AI would already be
234
00:14:55.400 --> 00:14:56.233
showing up.
235
00:14:56.760 --> 00:15:00.860
If AI had the capability at translation as it did at software engineering,
236
00:15:01.650 --> 00:15:02.220
naively,
237
00:15:02.220 --> 00:15:05.710
I might kind of feel like software engineers should be all going unemployed or
238
00:15:05.750 --> 00:15:08.810
something, and I don't think AI's quite there yet.
239
00:15:09.530 --> 00:15:11.890
And so if we're not seeing it with call center workers,
240
00:15:11.950 --> 00:15:14.770
that makes me feel like there's just other stuff going on.
241
00:15:15.990 --> 00:15:20.950
And then last one, energy as the main bottleneck of AI, R&D and deployment.
242
00:15:21.190 --> 00:15:22.023
Peter.
243
00:15:22.530 --> 00:15:27.130
I think it's underrated as an aspect of how AI is already reshaping the economy,
244
00:15:27.210 --> 00:15:31.470
the impact on factor markets as being both a bottleneck to deployment,
245
00:15:31.670 --> 00:15:32.503
but also
246
00:15:34.270 --> 00:15:38.950
the impact of this rapid CapEx on interest rates and investment
247
00:15:39.030 --> 00:15:41.670
elsewhere in the economy is a material effect.
248
00:15:42.770 --> 00:15:46.590
Underrated to the extent that the policy environment is not prioritizing
249
00:15:46.630 --> 00:15:47.790
bringing new energy online.
250
00:15:48.950 --> 00:15:50.790
While agreeing with everything that was just said,
251
00:15:50.890 --> 00:15:55.750
I'll say overrated in the sense that supply is elastic in a lot
252
00:15:55.810 --> 00:15:56.970
more ways than people expect.
253
00:15:57.250 --> 00:16:02.190
We'll find ways to get energy even if people need to become electricians and we
254
00:16:02.230 --> 00:16:03.530
need to pay them lots of money to do so.
255
00:16:05.350 --> 00:16:07.390
So I want to go back to John's talk this morning.
256
00:16:07.550 --> 00:16:11.490
He talked about the Coasean reading of AI and firm structure.
257
00:16:11.950 --> 00:16:14.890
And so I wanted to explore that a little bit more with you three. Ara,
258
00:16:15.010 --> 00:16:17.450
I'll start with you. Sorry. Ara, I'll start with you.
259
00:16:17.670 --> 00:16:19.150
We'll go Peter and then we'll go Basil.
260
00:16:20.270 --> 00:16:23.350
Is AI just mostly an efficiency enhancer?
261
00:16:23.410 --> 00:16:27.110
Is it just going to be same organization, same structure,
262
00:16:27.430 --> 00:16:28.910
just everybody works faster?
263
00:16:29.530 --> 00:16:32.850
Or do you think it's going to fundamentally change how we structure
264
00:16:33.010 --> 00:16:34.970
organizations and build businesses?
265
00:16:37.080 --> 00:16:40.450
I think if you look at the organizations that have done a really good job
266
00:16:40.610 --> 00:16:45.210
implementing AI, it's not fundamentally changing structures. There are,
267
00:16:46.410 --> 00:16:47.890
I think actually somewhat intentionally,
268
00:16:49.510 --> 00:16:53.110
the labs are explicitly calling people's job titles like member of technical
269
00:16:53.150 --> 00:16:54.650
staff or member of finance staff,
270
00:16:55.350 --> 00:16:57.390
and firms are suggesting that everyone should be a builder.
271
00:16:57.630 --> 00:16:59.190
So there's a little bit of that happening.
272
00:17:02.210 --> 00:17:05.630
At the same time, I think that for most firms,
273
00:17:06.190 --> 00:17:11.050
as their competitors start to adopt AI and then also move further along the
274
00:17:11.110 --> 00:17:12.070
AI adoption curve,
275
00:17:12.310 --> 00:17:16.890
they're all going to have to compete to be just as productive as each other.
276
00:17:17.460 --> 00:17:21.390
And AI just becomes this additional tool that
277
00:17:22.650 --> 00:17:26.330
makes everyone more productive, makes better products, makes better software,
278
00:17:27.090 --> 00:17:29.570
makes people a little bit more organized, but everyone has it.
279
00:17:30.740 --> 00:17:33.590
And in that kind of model,
280
00:17:36.590 --> 00:17:40.030
there will be shifts in the kinds of jobs that grow and fall,
281
00:17:42.960 --> 00:17:43.793
but
282
00:17:45.710 --> 00:17:48.870
the main outputs of an organization don't end up changing all that much.
They
283
00:17:48.910 --> 00:17:49.960
just become higher quality.
284
00:17:51.770 --> 00:17:55.010
So I guess I would say two things here. One, which is more qualitative,
285
00:17:55.070 --> 00:17:57.950
we did this large-scale survey of Claude users,
286
00:17:58.010 --> 00:18:01.290
about 81,000 people around the world. And in general,
287
00:18:01.350 --> 00:18:04.670
people said that Claude helped them become more productive.
288
00:18:05.270 --> 00:18:08.870
But one of the main things that they emphasized was not how fast you do
289
00:18:08.990 --> 00:18:13.350
something, but the scope of what you're able to do. So product managers,
290
00:18:13.430 --> 00:18:17.610
for example, being able to do some type of software engineering and vice versa,
291
00:18:17.750 --> 00:18:21.870
this blurring of the boundary between traditional job functions,
292
00:18:21.930 --> 00:18:26.170
I think is in some sense underway as you have broader scope.
293
00:18:26.810 --> 00:18:28.190
But at an organizational level,
294
00:18:28.310 --> 00:18:31.910
I think we should not downplay the importance of complimentary investments that
295
00:18:31.950 --> 00:18:35.310
businesses need to make in order to, at scale,
296
00:18:35.570 --> 00:18:37.210
generate the productivity benefits.
297
00:18:37.930 --> 00:18:42.650
So when we look at enterprise API deployment of Claude where Claude is embedded
298
00:18:42.710 --> 00:18:46.550
in existing or new workflows, the most complex tasks,
299
00:18:46.690 --> 00:18:48.950
something like automated biological research,
300
00:18:49.510 --> 00:18:51.470
typically generates a bunch of tokens,
301
00:18:51.870 --> 00:18:56.850
but relies on disproportionately more input tokens or context information.
302
00:18:58.570 --> 00:19:01.430
And that pattern holds quite strongly.
303
00:19:02.650 --> 00:19:05.870
What that reveals is that it might not just be capabilities alone,
304
00:19:05.930 --> 00:19:09.490
but access to the relevant information at the right time.
305
00:19:09.610 --> 00:19:12.510
And I think about what tasks might Claude be good at today,
306
00:19:12.650 --> 00:19:15.270
but where it's very hard to get the right information.
307
00:19:15.630 --> 00:19:19.350
If you're developing a sales strategy and the information that Claude needs is
308
00:19:19.390 --> 00:19:20.570
in your coworker's mind,
309
00:19:21.650 --> 00:19:25.790
you need to think about organizational processes to elicit that information or
310
00:19:25.850 --> 00:19:29.490
data modernization to connect all the data pieces together.
311
00:19:29.890 --> 00:19:33.210
I think that points in the direction of sort of economies of scope.
312
00:19:33.550 --> 00:19:34.710
Maybe in equilibrium,
313
00:19:35.090 --> 00:19:39.130
firms will have a huge incentive to broaden out the range of things that they do
314
00:19:39.650 --> 00:19:43.830
so that they can codify and centralize that information. On the other hand,
315
00:19:44.150 --> 00:19:48.250
creative destruction is an incredible force for productivity growth and economic
316
00:19:48.290 --> 00:19:48.630
growth.
317
00:19:48.630 --> 00:19:53.470
And so the barriers to entry as a startup are now much lower.
318
00:19:53.670 --> 00:19:56.250
You can use Claude to broaden the scope of what you do.
319
00:19:56.330 --> 00:19:59.390
And I think how this all shakes out, a little unclear.
320
00:20:00.830 --> 00:20:03.110
Just to, I think, echo the points that were made.
321
00:20:03.270 --> 00:20:07.350
So one way to think about one of the main economic effects of the web was
322
00:20:07.430 --> 00:20:09.710
sending the cost of communication nearly to zero,
323
00:20:10.230 --> 00:20:12.830
and that had clear effects on making remote work possible,
324
00:20:13.210 --> 00:20:17.170
making email maybe the main form of communication within firms.
325
00:20:17.670 --> 00:20:20.250
And so what cost will AI change?
326
00:20:20.550 --> 00:20:22.850
I think it is something about idea generation,
327
00:20:23.470 --> 00:20:26.830
the cost of that getting sent down to zero and how that changes the scope of
328
00:20:26.890 --> 00:20:29.990
activities of any individual worker or the firm as a whole.
329
00:20:31.370 --> 00:20:34.370
So I want to make sure we get in questions about the labor market,
330
00:20:34.690 --> 00:20:38.110
because there's a lot of anxiety about how AI is affecting that.
331
00:20:38.810 --> 00:20:40.050
So in particular,
332
00:20:40.230 --> 00:20:44.910
I think that there is an anxiety around younger workers and new hires. So Basil,
333
00:20:44.970 --> 00:20:47.030
I'll start with you and we'll come down the line.
334
00:20:47.870 --> 00:20:52.590
To what extent do we see actual displacement of any worker,
335
00:20:52.650 --> 00:20:54.830
but in particular young workers from AI?
336
00:20:56.030 --> 00:20:57.830
And if you want to opine on this,
337
00:20:59.370 --> 00:21:01.590
if we do see displacement effects for young workers,
338
00:21:01.650 --> 00:21:04.770
how does that affect things like apprenticeships, learning, new skills?
339
00:21:05.030 --> 00:21:09.450
How do we upskill young workers in an environment where AI is displacing them?
340
00:21:10.130 --> 00:21:13.570
This is the $30 trillion question and a very difficult one,
341
00:21:13.690 --> 00:21:17.010
and there's research with evidence all over the place.
342
00:21:17.210 --> 00:21:22.080
So I'll just focus on one point, which is that ChatGPT came out November 2022.
343
00:21:22.750 --> 00:21:27.510
This is the same time the Federal Reserve in the United States or really began
344
00:21:27.650 --> 00:21:29.390
hiking interest rates to fight inflation.
345
00:21:30.030 --> 00:21:34.530
Higher interest rates change the economic calculus of a firm. If I'm a firm,
346
00:21:34.610 --> 00:21:36.090
if I'm a company, I'm trying to decide
347
00:21:38.250 --> 00:21:40.290
whether or not to invest in something for the long term,
348
00:21:40.710 --> 00:21:43.390
higher interest rates make it less likely that I'll want to invest.
349
00:21:44.390 --> 00:21:47.250
Hiring a young worker is like making a long-term investment.
350
00:21:47.670 --> 00:21:49.470
You're making a long-term investment in an individual.
351
00:21:49.970 --> 00:21:54.950
So you might imagine that the rise in interest rates has caused a tighter labor
352
00:21:55.010 --> 00:21:56.570
market for younger workers,
353
00:21:56.930 --> 00:22:00.390
and it's really hard to disentangle that effect from the effect of AI on
354
00:22:00.430 --> 00:22:01.890
workers.
How it nets out,
355
00:22:02.410 --> 00:22:05.950
my read is that it's unclear maybe there's some effect,
356
00:22:06.050 --> 00:22:08.450
but mostly it's these other macro conditions.
357
00:22:11.150 --> 00:22:13.190
I think that the labor market effects are going to start with the most
358
00:22:13.250 --> 00:22:14.370
marginally attached workers.
359
00:22:14.430 --> 00:22:19.250
We saw that in Ramp data where we ran a study on firms that had previously
360
00:22:19.350 --> 00:22:22.390
spent a sizable amount of revenue on labor marketplaces,
361
00:22:22.510 --> 00:22:24.590
so Upwork and Fiverr,
362
00:22:26.370 --> 00:22:28.430
tasks that are generally very well scoped,
363
00:22:29.470 --> 00:22:32.150
tasks that are very specific and tasks that are very short term.
364
00:22:32.790 --> 00:22:33.890
That can include design tasks,
365
00:22:33.950 --> 00:22:36.510
but it can also include finance tasks and software engineering tasks.
366
00:22:36.910 --> 00:22:40.650
So firms that used to spend a lot of money on these freelance labor marketplaces
367
00:22:41.330 --> 00:22:45.790
have since cut their spending and shifted it over to
368
00:22:46.250 --> 00:22:51.200
AI models, such that we're seeing something between savings of 60% to 97%
369
00:22:52.550 --> 00:22:55.310
amongst the firms that used to spend a lot of money on those labor marketplaces.
370
00:22:55.950 --> 00:22:58.890
And the firms that made that shift fastest were those that were most
371
00:22:59.050 --> 00:23:00.850
economically incentivized to do so,
372
00:23:00.970 --> 00:23:05.770
the ones that spent the highest share of their revenue on labor marketplaces in
373
00:23:05.790 --> 00:23:10.210
the first place.
And so not only is there this economic incentive for firms and
374
00:23:10.250 --> 00:23:11.430
now this opportunity to
375
00:23:13.810 --> 00:23:18.810
move work back in house and maybe make your own logo or design
376
00:23:18.850 --> 00:23:21.630
or quick marketing copy if you need to instead of outsourcing it.
377
00:23:23.050 --> 00:23:27.330
But the workers that are most exposed to that are not captured by
378
00:23:28.470 --> 00:23:31.810
the research that's currently happening on the economic impact of AI.
379
00:23:32.290 --> 00:23:36.050
When we see these papers that are written about AI exposure,
380
00:23:36.110 --> 00:23:39.470
they're talking about jobs that are very clearly defined,
381
00:23:39.530 --> 00:23:40.950
but they're not talking about freelance workers.
382
00:23:41.890 --> 00:23:45.610
Freelance workers are not often well captured in government data sets anyway.
383
00:23:47.010 --> 00:23:51.330
And also they're exactly the kind of worker for whom our current unemployment
384
00:23:51.370 --> 00:23:53.370
system is not designed to
385
00:23:55.010 --> 00:23:59.950
serve in the event of some large scale unemployment issue.
386
00:24:00.790 --> 00:24:04.110
And so that's one of the places where we are going to see one of the early
387
00:24:04.150 --> 00:24:04.890
impacts of AI,
388
00:24:04.890 --> 00:24:09.510
where we are already seeing that amongst the firms that are most incentivized to
389
00:24:09.550 --> 00:24:13.760
make that change and where we don't have policies in place to
390
00:24:17.870 --> 00:24:19.950
absorb that negative effect.
391
00:24:22.950 --> 00:24:25.910
I'll kind of return to this point that Basil made, which is,
392
00:24:26.010 --> 00:24:30.870
we had the largest nonrecessionary labor market slowdown in US
393
00:24:30.930 --> 00:24:31.763
history.
394
00:24:32.170 --> 00:24:36.770
Who typically struggles when the economy slows down?
395
00:24:36.990 --> 00:24:40.670
It is those who are just entering the labor force.
396
00:24:41.350 --> 00:24:44.910
I think you have to have an additional piece of the puzzle to argue that it's
397
00:24:45.070 --> 00:24:45.750
that.
398
00:24:45.750 --> 00:24:50.730
It has to be that young workers are more cyclically sensitive in occupations
399
00:24:50.770 --> 00:24:55.690
that have higher AI exposure. That's like my instinct.
400
00:24:55.830 --> 00:24:59.610
We had some... I mean, unfortunately, in some sense,
401
00:24:59.690 --> 00:25:03.830
we had the arrival of this incredibly consequential technology in a very
402
00:25:04.110 --> 00:25:06.270
volatile macroeconomic environment,
403
00:25:06.330 --> 00:25:10.030
which makes it incredibly challenging to tease out its effect from other
404
00:25:10.070 --> 00:25:10.903
effects.
405
00:25:11.710 --> 00:25:15.030
That was the motivation for this report that we put out in early March where we
406
00:25:15.110 --> 00:25:19.530
asked the question, "If displacement effects materialize,
407
00:25:19.870 --> 00:25:24.610
where might we expect them to show up at the very least?" So it's a very weak
408
00:25:24.910 --> 00:25:26.830
test, and that generated this,
409
00:25:27.890 --> 00:25:32.490
where's Claude being used for tasks and automated ways for work purposes.
410
00:25:33.290 --> 00:25:34.030
By that measure,
411
00:25:34.030 --> 00:25:39.010
there's no systematic movement of unemployment for workers in the most
412
00:25:39.130 --> 00:25:44.030
observed exposure relative to those in jobs that have no AI
413
00:25:44.090 --> 00:25:47.190
exposure. But when we look at younger workers,
414
00:25:47.330 --> 00:25:52.310
it does seem like job-finding rates are a little bit lower in the
415
00:25:52.370 --> 00:25:55.510
last year or so for those most exposed.
416
00:25:56.650 --> 00:25:57.790
In the survey that we ran,
417
00:25:58.130 --> 00:26:02.630
we also document that the people who are most concerned about the threat
418
00:26:03.070 --> 00:26:07.970
to their job from AI are exactly those workers in occupations that have this
419
00:26:08.050 --> 00:26:11.730
high-AI exposure as we measure it. So at the very least,
420
00:26:11.790 --> 00:26:15.570
people are worried about displacement in exactly the places that we might expect
421
00:26:15.630 --> 00:26:17.050
it to materialize.
422
00:26:17.170 --> 00:26:21.730
And then young workers in particular are most worried among the people that we
423
00:26:21.990 --> 00:26:22.823
surveyed.
424
00:26:23.270 --> 00:26:25.470
Yeah. I often say in macroeconomics,
425
00:26:25.730 --> 00:26:27.990
you don't have the chance to run a lot of experiments.
426
00:26:28.190 --> 00:26:31.590
You can't put a country in recession and compare it to another country,
427
00:26:31.710 --> 00:26:36.310
so you have to accept natural experiments. And as exciting as AI is,
428
00:26:36.510 --> 00:26:39.970
one of the most unfortunate things about the timing of AI,
429
00:26:40.030 --> 00:26:42.120
going back to Basil's point, is that it happened...
430
00:26:42.710 --> 00:26:46.250
We got ChatGPT the same year that we got a hike in interest rates,
431
00:26:47.030 --> 00:26:50.730
that we got the great resignation, living in the shadow of the pandemic.
432
00:26:51.350 --> 00:26:55.310
And more recently, there have been all these other shocks affecting the economy,
433
00:26:55.590 --> 00:26:58.370
tariffs, immigration policies.
434
00:26:59.030 --> 00:27:01.510
It just makes it really hard to disentangle.
435
00:27:01.650 --> 00:27:06.050
I really wish there were a cleaner way to identify what AI is doing.
436
00:27:08.650 --> 00:27:11.570
We only have three minutes left. So let me end on this question.
437
00:27:12.070 --> 00:27:16.710
What is one thing that most people currently believe about AI and the economy
438
00:27:17.110 --> 00:27:21.490
that you think will look obviously wrong in two years? And Peter,
439
00:27:21.550 --> 00:27:22.383
I'll start with you.
440
00:27:24.950 --> 00:27:25.750
Obviously wrong.
441
00:27:25.750 --> 00:27:30.030
I'll focus on something that I think people need to be paying more attention to,
442
00:27:30.590 --> 00:27:33.970
which is in the long run, the big thing that matters is,
443
00:27:34.370 --> 00:27:37.850
does AI automate the process of innovation itself?
444
00:27:38.310 --> 00:27:41.550
You write down your favorite growth model of the economy.
445
00:27:42.270 --> 00:27:45.050
Productivity growth is the engine of long run prosperity,
446
00:27:45.570 --> 00:27:49.910
and AI has this possibility of automating innovation,
447
00:27:50.050 --> 00:27:53.030
overcoming the burden of knowledge, helping us,
448
00:27:53.350 --> 00:27:55.110
and innovation and the method of innovation.
449
00:27:56.030 --> 00:27:58.170
That's a very urgent thing for us to understand.
450
00:27:58.390 --> 00:28:03.130
Labs sometimes talk about this in the context of does AI help
451
00:28:04.690 --> 00:28:06.570
accelerate AI innovation itself,
452
00:28:06.630 --> 00:28:09.990
but it's actually a broader consideration for the economy. All.
453
00:28:10.410 --> 00:28:13.310
Right. I'm going to do something no one ever does.
454
00:28:13.370 --> 00:28:16.300
I'm going to say something bad about Anthropic and...
455
00:28:17.570 --> 00:28:18.403
Love it.
456
00:28:21.130 --> 00:28:25.930
The labs are not incentivized to produce products that
457
00:28:26.110 --> 00:28:29.930
are effective and priced effectively.
458
00:28:30.070 --> 00:28:33.250
You are incentivized to spend as much tokens as possible on these kinds of
459
00:28:33.290 --> 00:28:34.123
products and tools.
460
00:28:34.370 --> 00:28:37.210
And I do think that's going to start to make its way into business
461
00:28:37.250 --> 00:28:38.083
decision-making,
462
00:28:38.690 --> 00:28:43.630
where companies may opt for software and tools that are not explicitly built
463
00:28:43.710 --> 00:28:48.550
by the labs because a competitor is better incentivized to build
464
00:28:48.610 --> 00:28:52.390
something that is effective and priced effectively and uses tokens in a
465
00:28:52.450 --> 00:28:53.283
moderated way.
466
00:28:54.270 --> 00:28:57.090
And the second point being that that would also have employment effects,
467
00:28:57.530 --> 00:29:02.130
that if the labs are not able to control the
468
00:29:02.230 --> 00:29:03.470
rising cost of tokens,
469
00:29:04.230 --> 00:29:06.970
because we're seeing all these stories about large companies like Uber blowing
470
00:29:06.990 --> 00:29:07.890
through their token budgets,
471
00:29:08.590 --> 00:29:13.370
then the cost benefit analysis of adding labor versus
472
00:29:13.550 --> 00:29:15.230
adding AI tokens completely falls apart.
473
00:29:17.390 --> 00:29:19.970
I'll give a take that won't be resolved on a two-year time horizon,
474
00:29:20.090 --> 00:29:23.210
but I think speaks to something that you hear a lot about,
475
00:29:23.710 --> 00:29:27.110
which is the idea of escaping the permanent underclass. Bad news,
476
00:29:27.210 --> 00:29:28.010
I don't think it's possible,
477
00:29:28.010 --> 00:29:29.450
you're not going to escape the permanent underclass.
478
00:29:29.830 --> 00:29:33.930
What I mean specifically is that either AI is like a normal technology,
479
00:29:34.010 --> 00:29:36.270
it's the web on steroids, happens five times faster,
480
00:29:36.810 --> 00:29:38.630
and everything's hunky-dory.
481
00:29:38.770 --> 00:29:42.050
We have normal political system for thinking about redistribution, et cetera,
482
00:29:42.270 --> 00:29:45.530
despite all its problems, or the world becomes totally sci-fi.
483
00:29:45.850 --> 00:29:48.990
We're all in the permanent underclass. Saving now is not doing you any good.
484
00:29:49.170 --> 00:29:53.870
It's like being the Incas before Columbus comes to Americas and trying to save
485
00:29:53.890 --> 00:29:56.070
your way out of the effects of colonization.
486
00:29:57.790 --> 00:30:01.030
Either that happens or the politicals system figures it out.
487
00:30:01.590 --> 00:30:02.970
We need the political system to figure it out.
488
00:30:04.430 --> 00:30:05.760
That was a fantastic conversation.
489
00:30:05.820 --> 00:30:09.460
So please join me in a round of applause for our panelists.